Date of Award

Spring 5-8-2023

Access Type

Dissertation - Open Access

Degree Name

Doctor of Philosophy in Aerospace Engineering

Department

Aerospace Engineering

Committee Chair

Dr. Hever Moncayo

Committee Advisor

Dr. Hever Moncayo

First Committee Member

Dr. Kadriye Merve Dogan

Second Committee Member

Dr. Troy Henderson

Third Committee Member

Dr. Richard Stansbury

Fourth Committee Member

Dr. Maj Mirmirani

College Dean

Dr. James W. Gregory

Abstract

The increasing complexity and autonomy of modern systems, particularly in the aerospace industry, demand robust and adaptive fault detection and health management solutions. The development of a data-driven fault detection system that can adapt to varying conditions and system changes is critical to the performance, safety, and reliability of these systems. This dissertation presents a novel fault detection approach based on the integration of the artificial immune system (AIS) paradigm and Online Support Vector Machines (OSVM). Together, these algorithms create the Artificial Immune System augemented Online Support Vector Machine (AISOSVM).

The AISOSVM framework combines the strengths of the AIS and OSVM to create a fault detection system that can effectively identify faults in complex systems while maintaining adaptability. The framework is designed using Model-Based Systems Engineering (MBSE) principles, employing the Capella tool and the Arcadia methodology to develop a structured, integrated approach for the design and deployment of the data-driven fault detection system. A key contribution of this research is the development of a Clonal Selection Algorithm that optimizes the OSVM hyperparameters and the V-Detector algorithm parameters, resulting in a more effective fault detection solution. The integration of the AIS in the training process enables the generation of synthetic abnormal data, mitigating the need for engineers to gather large amounts of failure data, which can be impractical.

The AISOSVM also incorporates incremental learning and decremental unlearning for the Online Support Vector Machine, allowing the system to adapt online using lightweight computational processes. This capability significantly improves the efficiency of fault detection systems, eliminating the need for offline retraining and redeployment.

Reinforcement Learning (RL) is proposed as a promising future direction for the AISOSVM, as it can help autonomously adapt the system performance in near real-time, further mitigating the need for acquiring large amounts of system data for training, and improving the efficiency of the adaptation process by intelligently selecting the best samples to learn from.

The AISOSVM framework was applied to real-world scenarios and platform models, demonstrating its effectiveness and adaptability in various use cases. The combination of the AIS and OSVM, along with the online learning and RL integration, provides a robust and adaptive solution for fault detection and health management in complex autonomous systems.

This dissertation presents a significant contribution to the field of fault detection and health management by integrating the artificial immune system paradigm with Online Support Vector Machines, developing a structured, integrated approach for designing and deploying data-driven fault detection systems, and implementing reinforcement learning for online, autonomous adaptation of fault management systems. The AISOSVM framework offers a promising solution to address the challenges of fault detection in complex, autonomous systems, with potential applications in a wide range of industries beyond aerospace.

Download

Share

COinS